AI in Health: Genomics Breakthroughs, Ad Bias Risks, and Radiology Gains

This week in health AI brought advances across genomics, cardiology, radiology, and drug discovery—alongside a sobering reminder that systems we trust with clinical decisions can be influenced in ways we don't always see.

AlphaGenome 2 Predicts How DNA Variants Drive Disease in Specific Tissues

Google DeepMind released AlphaGenome 2, a significant upgrade to its AI model for predicting the functional consequences of genetic variants at single-cell resolution. Trained on more than 107 million cell profiles, the model can now identify how a given DNA change alters gene regulation in specific tissue types—a capability that matters enormously when the same variant may be benign in one cell context and disease-driving in another.

The benchmark numbers are striking: AlphaGenome 2 achieves 30% causal variant recovery at 90% precision, placing it well ahead of prior approaches. For researchers studying complex diseases like type 2 diabetes, rheumatoid arthritis, or heart failure—where genome-wide association studies have long produced hits that resist mechanistic interpretation—this kind of tissue-aware variant scoring could meaningfully narrow down which signals to pursue experimentally.

By operating at cell-type resolution rather than bulk tissue averages, the model captures the specificity that makes human disease so difficult to model: the same gene can behave very differently in a pancreatic beta cell versus a macrophage. It is part of DeepMind's broader push into genomic foundation models and builds on the original AlphaGenome architecture.

Drug Ads Can Shift AI Treatment Recommendations by Nearly 50 Percentage Points

A study published in Lancet Digital Health by researchers at Mount Sinai delivered one of the more unsettling findings in clinical AI this year. The team tested whether pharmaceutical advertisements embedded in context—the kind of promotional material that fills medical websites and apps—could influence which treatments AI chatbots recommended to simulated patients.

They could. Across a range of clinical scenarios, AI recommendations for the advertised drug rose from a baseline of 35% to 47.6%—a shift of nearly 13 percentage points on average. In four specific scenarios, the effect was more extreme: models went from recommending the advertised drug 0% of the time to 100% once ads were introduced.

This matters because AI chatbots are increasingly embedded in patient-facing health tools, and because the mechanisms by which large language models absorb and act on promotional content are not transparent. Unlike a physician who might consciously recognize a pharmaceutical sales pitch, an AI processes it as text—and that text, apparently, shapes outputs in ways that could favor commercial interests over clinical ones. The study is a call for systematic auditing of AI health tools deployed in environments where advertising is present.

Oxford JAMA Cardiology Study Validates AI Cardiac Inflammation Score in Nearly 20,000 Patients

Caristo Diagnostics announced the publication of a landmark Oxford-led study in JAMA Cardiology validating its CaRi-Heart technology in 19,110 patients. The AI-derived score measures coronary inflammation from routine CT scans—information that has historically been invisible to standard imaging reads.

The study's headline finding is that CaRi-Heart reclassifies nearly half of cardiovascular deaths that occur in patients with a coronary artery calcium (CAC) score of zero. CAC scoring is widely used to stratify cardiac risk, and a score of zero is often interpreted as reassuring. But roughly half of heart attacks occur in people without significant calcification—the so-called zero-CAC paradox. An AI tool that can flag coronary inflammation in these patients, before calcification develops, would address a genuine clinical blind spot.

The validation in a cohort of nearly 20,000 patients gives the finding statistical weight that smaller feasibility studies could not. CaRi-Heart joins a growing list of AI-derived biomarkers demonstrating additive prognostic value beyond established risk metrics.

AI Visual Language Models Improve Radiology Detection by Up to 57% at RSNA 2026

Mosaic Clinical Technologies presented data at RSNA 2026 showing that its AI visual language models (VLMs) achieved an average 40–43% improvement in detection rates across 50 clinical conditions. For lung nodules on chest X-ray—a notoriously challenging and high-stakes detection task—the improvement reached 57.5%.

The company also reported that AI-drafted radiology reports were comparable in accuracy to reports written by radiologists, a finding that will draw scrutiny but also attention from health systems facing radiologist shortages. VLMs differ from earlier radiology AI in that they generate natural-language output alongside detection, rather than producing only bounding boxes or probability scores. That capability makes them more directly useful as report-drafting tools.

RSNA 2026 saw a broader theme of AI moving from detection assistance to integrated workflow tools, with multiple vendors demonstrating systems that handle the full arc from image ingestion to draft report generation. Mosaic's data represents one of the more specific quantitative claims made at the conference.

Sino Biological's 4RISE Platform Bridges AI Molecule Design and Wet Lab Validation

Sino Biological launched 4RISE, an integrated platform designed to close the gap between computationally designed molecules and experimental wet-lab validation. The platform supports antibodies, enzymes, and other protein candidates generated by AI design tools, providing the assays and biological testing infrastructure needed to confirm whether a computationally promising molecule actually behaves as predicted in biological systems.

The launch addresses a persistent bottleneck in AI-driven drug discovery: AI tools have become increasingly capable of proposing novel molecular structures, but the experimental infrastructure to rapidly validate those proposals has lagged. 4RISE aims to serve biotech companies and academic labs running AI-first discovery programs who need a bridge from in silico output to biological data without building full wet lab capability internally. The platform spans antibody and enzyme validation workflows and is designed to integrate with upstream AI design pipelines, positioning Sino Biological as an infrastructure provider for the AI drug discovery ecosystem.